Learning Commute-Time-Preserving World Models for Planning
The paper tackles the problem of learning task-agnostic world models for planning in latent space. It introduces CTWM, a reconstruction-free method that combines residual prediction with a log-determinant regularizer to recover the scaled Laplacian representation, so that latent L2 distances correspond to environment commute-time distances. This matters because it enables zero-shot goal-reaching without reward, and empirically CTWM matches or beats the LeWM baseline while using half the parameters.
CTWM is a well-motivated contribution that bridges spectral graph theory and joint-embedding predictive architectures for model-based RL. The theoretical result cleanly identifies why isotropic collapse-prevention harms commute-time geometry, and the logdet fix is elegant. Empirically, CTWM improves over LeWM on several continuous control and navigation tasks. However, the theoretical guarantee relies on deterministic, reversible dynamics and a predictor fixed-point assumption, which limits its applicability to general stochastic environments. Additionally, the strongest validation of the exact scaling is limited to a synthetic discrete torus, and the authors disclose that generative AI assisted with the proof in Appendix C.3.
The theoretical analysis is the paper’s anchor. Theorem 1 and its corollaries show that the logdet regularizer yields the correct \lambda_k^{-1/2} scaling, unlike orthonormality or VCReg heuristics that enforce isotropy. On the discrete torus, the learned variance spectrum tracks the ground-truth Laplacian (Fig. 2c) and squared L2 distances correlate with commute times (R^2=0.95). In continuous domains, CTWM consistently outperforms LeWM while using half the parameters, with especially large margins on Cube, Scene, and Pointmaze.
The recovery guarantee requires deterministic and reversible dynamics, yet the continuous benchmarks are stochastic. The proof also assumes the predictor has reached its fixed point, which is not enforced during training. The only direct empirical verification of exact scaling is on a synthetic discrete torus; in continuous environments the authors resort to a discretized Pointmaze approximation.
Planning gains are attributed to commute-time geometry, but CTWM changes multiple implementation details relative to LeWM—layer norm replaces batch norm, weight decay is removed, and the latent dimension is halved—making it hard to isolate the effect of the logdet regularizer. Finally, several experimental datasets and the LeWM baseline itself are drawn from as-yet-unpublished 2026 works.
In the discrete domain the evidence is strong: CTWM is the only method that tracks the \lambda_k^{-1/2} variance scaling and attains R^2=0.95 against ground-truth commute times, whereas VICReg and WM-GDO flatten the spectrum. The comparison to LeWM in continuous domains is favorable, but the paper does not empirically test whether other isotropic SSL methods (e.g., BYOL, Barlow Twins) would also underperform, or whether the gains are specific to SIGReg. The claim that isotropic regularizers degrade commute-time scaling is supported by prior art: VICReg explicitly decorrelates variables and enforces a variance floor, which biases toward isotropy.
The authors share code and report all hyperparameters in Appendix A, which aids reproduction. However, multiple benchmarks rely on datasets from future or very recent works (e.g., Maes et al., 2026a; Park et al., 2025), some of which may not yet be publicly available. The proof relies on the predictor fixed-point assumption; reproducing the theoretical result exactly would require verifying deterministic dynamics, which is not feasible on standard continuous control tasks. Additionally, AI tools were used to assist with the proof in Appendix C.3, which adds a small burden of independent verification.
World models allow agents to plan in latent space by choosing a sequence of actions that most reduces the distance to a given goal state. Thus, planning can benefit from latent representations whose distances mirror commute-times in the environment. The spectral embedding space of the graph Laplacian provides such a representation, if it obeys a specific eigenvalue-dependent scaling. Unfortunately, instantiating the graph Laplacian is intractable in large, continuous environments. Self-supervised learning offers a natural route to such commute-time-preserving embeddings at scale. However, here we show that existing methods, which commonly encourage isotropic representations to prevent representational collapse, tend to degrade the "correct" eigenvalue-dependent scaling, leading to an inaccurate representation of commute times. To address this problem, we introduce Commute-Time-Preserving World Models (CTWMs), combining a latent displacement predictor and a log-determinant regularizer that prevents collapse, which provably recover the correctly scaled Laplacian representation under reversible deterministic dynamics and at the predictor's fixed point. In numerical simulations, CTWM matches or outperforms LeWM, a task-agnostic baseline, on several complex, continuous goal-reaching benchmarks, while using half the parameters.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.